From Input-handling Flaws to Crashables: Security ZLessons From LLM-based Coding Tools

TOPIC: From Input-handling Flaws to Crashables: Security ZLessons From LLM-based Coding Tools ABSTRACT: Claude Code illustrates how LLM-based coding tools expand the attack surface. Design choices around approvals, parsing, and error handling can turn into security flaws. We present specific findings Kodem uncovered in Claude Code. Both issues highlight how LLM-based coding tools introduce new misconfiguration and input-handling risks. This talk dissects the issues, their broader implications for AI developer tools, and practical mitigations.